Estimation of detection/classification performance in sonar imagery using textural features and self-organising maps

Accurate estimation of detection/classification performance for sidescan sonar systems in Mine Counter-Measure (MCM) applications is important for informing mission tactics and adapting autonomous behaviors. The approach presented in this paper assumes that detection/classification performance can be estimated solely from historical data collected from similar surveys. This paper introduces an algorithm for comparing the textural difference of two surveys, and calculating a performance estimate from the most similar surveys in a database. The approach characterizes surveys using a Self-Organizing Map (SOM) of textural features. A histogram is generated for each mission by assigning textures to their Best Matching Unit (BMU) in the SOM. Mission difference is calculated as the chi-squared distance between two histograms. Using this difference measure, it is shown that the estimate of detector recall over the most textually similar surveys improves on the estimate over all surveys by a factor of 5. The estimate of precision is improved by a factor of 2. For a deep learning based detector the mean error in the recall estimate is reduced to 2 +/- 2%.